R Deep Learning Essentials - Second Edition

The array.batch.size parameter

We only had 400 instances (rows) in our data, which can easily fit into memory. However, if your input data has millions of instances, the data needs to be split into batches during training in order to fit in the memory of the CPU/GPU. The number of instances you train at a time is the batch size. Note, you still iterate over all the data for the number of epochs, you just split the data into batches during each iteration and run the forward-propagation, backpropagation step over each batch for each epoch. For example, if you had 100 instances and selected a batch size of 32 with 6 epochs, you would need 4 batches for each epoch (100/32 = 3.125, so we need 4 batches to process all the data), for a total of ...